[CPU] [Snippets] Support inference precision f16 on Snippets for ARM

openvinotoolkit · Sep 9, 2024 · 536086c · 536086c
1 parent 4643e5f
commit 536086c
Show file tree

Hide file tree

Showing 3 changed files with 4 additions and 12 deletions.
diff --git a/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp b/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
@@ -925,13 +925,7 @@ void Transformations::MainSnippets(void) {
         CPU_REGISTER_PASS_X64(snippetsManager, SnippetsMarkSkipped, inferencePrecision == ov::element::bf16);
 #endif
     }
-    CPU_REGISTER_PASS_X64(snippetsManager, snippets::pass::SnippetsTokenization, tokenization_config);
-    // [126738] Remove precision constraint when Convert emitters are implemented on arm platform
-    // The redundant "if defined", used to WA error of "empty controlled statement found" should also be removed then.
-#if defined(OPENVINO_ARCH_ARM64)
-    if (inferencePrecision == ov::element::f32)
-        CPU_REGISTER_PASS_ARM(snippetsManager, snippets::pass::SnippetsTokenization, tokenization_config);
-#endif
+    CPU_REGISTER_PASS_COMMON(snippetsManager, snippets::pass::SnippetsTokenization, tokenization_config);
 
     // - MHA has BRGEMM that is supported only on AVX512 platforms
     // - CPU Plugin Subgraph supports only f32, bf16 (and quantized) BRGEMM

diff --git a/src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp b/src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp
@@ -499,8 +499,6 @@ std::vector<std::string> disabledTestPatterns() {
     retVector.emplace_back(R"(smoke_Snippets_Eltwise/TwoInputsAndOutputs.*)");
     // arm jit_eltwise_emitters doesn't support jit_power_dynamic_emitter yet
     retVector.emplace_back(R"(smoke_Snippets_Eltwise/MaxNumParamsEltwise.*)");
-    // TODO: 141292
-    retVector.emplace_back(R"(smoke_Snippets_Eltwise_FP16.*)");
 #endif
 #if defined(_WIN32)
     retVector.emplace_back(R"(.*smoke_QuantizedConvolutionBatchNormTransposeOnWeights/QuantizedConvolutionBatchNorm.CompareWithRefs/conv_type=convolution_quantize_type=fake_quantize_intervals_type=per_(tensor|channel)_transpose_on_weights=true_device=CPU.*)");

diff --git a/src/plugins/intel_cpu/tests/functional/shared_tests_instances/snippets/add.cpp b/src/plugins/intel_cpu/tests/functional/shared_tests_instances/snippets/add.cpp
@@ -27,7 +27,7 @@ INSTANTIATE_TEST_SUITE_P(smoke_Snippets_Eltwise, Add,
                          ::testing::Combine(
                              ::testing::ValuesIn(inShapesStatic1),
                              ::testing::ValuesIn(inShapesStatic2),
-                             ::testing::Values(ov::element::f32),
+                             ::testing::ValuesIn({ov::element::f32, ov::element::f16}),
                              ::testing::Values(1), // Add
                              ::testing::Values(1), // Subgraph is created, since the inputs are followed by converts
                              ::testing::Values(ov::test::utils::DEVICE_CPU)),
@@ -50,7 +50,7 @@ INSTANTIATE_TEST_SUITE_P(smoke_Snippets_Eltwise_Add, Add,
                          ::testing::Combine(
                                  ::testing::ValuesIn(inShapesDynamic1),
                                  ::testing::ValuesIn(inShapesDynamic2),
-                                 ::testing::Values(ov::element::f32),
+                                 ::testing::ValuesIn({ov::element::f32, ov::element::f16}),
                                  ::testing::Values(1),
                                  ::testing::Values(1), // Subgraph is created, since the inputs are followed by converts
                                  ::testing::Values(ov::test::utils::DEVICE_CPU)),
@@ -86,7 +86,7 @@ std::vector<std::vector<InputShape>> inShapesAddPair {
 INSTANTIATE_TEST_SUITE_P(smoke_Snippets_Eltwise, AddPair,
                          ::testing::Combine(
                                  ::testing::ValuesIn(inShapesAddPair),
-                                 ::testing::Values(ov::element::f32),
+                                 ::testing::ValuesIn({ov::element::f32, ov::element::f16}),
                                  ::testing::Values(1),
                                  ::testing::Values(1), // Subgraph is created, since the inputs are followed by converts
                                  ::testing::Values(ov::test::utils::DEVICE_CPU)),